DeepSeek-V4-Flash-0731 Refusal Directions
A 757 KB On-the-Fly Censorship Kill Switch
2 min readupdated
Usually, removing censorship (abliteration — stripping refusal directions from model weights) means downloading a separate ~157 GB copy of the model and restarting the server every time you want to switch. This repo by wswzjbg solves the same problem for DeepSeek-V4-Flash-0731 using a 757 KB direction-vector file that sits right alongside the original checkpoint. You adjust censorship strength in real time via an HTTP request (the λ parameter): at λ=0, the model is bit-for-bit identical to the original, while at λ=1.5 it stops refusing prompts without a reload or any drop in quality. It is built for anyone running DeepSeek-V4-Flash on vLLM who wants flexible filter controls or clean A/B tests on a single instance.
Key features
- Tiny footprint: The
refusal_dirs.safetensorsfile is just 757 KB (46 unit vectors of dimension 4096 in float32) instead of a second full ~157 GB checkpoint. - Smooth tuning: Adjust λ on the fly from 0 (original model) to 1.5 (zero refusals), or even push it into negative values to make the model more cautious than stock.
- No quality degradation: Unlike baked abliterated checkpoints that overshoot the direction (~240%) and tank speed, runtime projection preserves DSpark acceptance and long-context performance.
- Bit-exact at λ=0: Original model weights remain untouched, so you can verify their sha256 checksum directly against the official DeepSeek release.
What you can do
- Toggle modes in production: Enable or disable refusal removal on a live vLLM server with a single POST request without dropping active sessions.
- Run clean A/B tests: Compare stock and uncensored behavior on the exact same hardware, process, and load without metric drift.
- Fine-tune the balance: Dial in an intermediate λ value or a negative coefficient for workflows that call for extra restraint.
Example
curl -XPOST localhost:8888/admin/refusal_lambda -d '{"lambda": 1.5}' # без цензуры
curl -XPOST localhost:8888/admin/refusal_lambda -d '{"lambda": 0}' # обратно в стокRead more
DetailsMetrics at λ=1.5: 0/10 refusals (vs. 9/10 stock), NIAH 30/30 at 32k+128k, tool-calling 8/8, DSpark acceptance 0.5608.
Benchmarked on 2× DGX Spark GB10 (TP=2, vLLM 0.25.2 / 0.27.1 DSpark, context up to 262k–393k).
Requires a vLLM patch from the linked GitHub repo; changing λ warms up a fresh prefix cache for the new value.
Vectors extracted via the Arditi et al. (NeurIPS 2024) method from the
cebeuq/DeepSeek-V4-Flash-0731-abliterated checkpoint.









Comments